TraceCDS answers from FDA drug labeling first, and a trust-checked web search when the labels don't cover a question. Every statement cites its source, every number is verified against that source, and directive language is removed — enforced in code, on every answer.
A hybrid retrieval and reranking pipeline grounds every open-source model response in a cited passage, then verifies the response before it's shown.
Images, patient identifiers, and time-critical questions are declined before any retrieval.
BM25 + BAAI/bge-m3 dense retrieval, reranked by BAAI/bge-reranker-v2-m3.
Extractive quotes, or a summary from a self-hosted open-source LLM (BioMistral).
Every statement is checked against its source; failures are removed and logged.
Prompting an LLM to follow FDA's non-device clinical decision support criteria isn't reliable on its own — these checks run in code, on every statement, regardless of which generator produced it.
Label section, DailyMed link, and label effective date — or a trust-checked web source.
A dose, frequency, or threshold that isn't in the cited source is removed.
"Give 500 mg" fails. "The label states the starting dose is 500 mg [S1]" passes.
Approved healthcare sources are marked; anything else is clearly flagged, never hidden.
When nothing can be verified against a source, TraceCDS says so instead of guessing.
Dosing tables are converted to Markdown and quoted row by row with their column headers.
Output from the synthetic test label in the project's test suite.
Q: How is exampledrug dosed for kidney function and eGFR?
No question is sent to an outside AI service. The model, the search engine, and the web app all run on infrastructure TraceCDS controls.
Free, self-hosted, and open source. Not a substitute for clinical judgment.
Use TraceCDS →